Multivariable prediction model for the need for surgery in horses with colic

A survey of 1,965 equine colic cases was conducted from August 1985 to July 1986 at 10 equine referral centers located throughout the United States. The purpose of this study was to develop and validate a multivariable model for the need for surgery. Two-thirds of the cases were randomly selected for model development (1,336), whereas the remaining cases (629) were used only for subsequent validation of the model. If a lesion requiring surgical correction was found at either surgery or necropsy, the case for the horse was classified as surgical, otherwise the case was classified as medical. Only variables that were significant (P < 0.05) in an initial bivariable screening procedure were considered in the model development. Because of the large number of missing values in the data set, only variables for which there were < 400 missing values were considered in the multivariable analysis. A multivariable logistic regression model was constructed by use of a stepwise algorithm. The model used 640 cases and included variables: rectal findings, signs of abdominal pain, peripheral pulse strength, and abdominal sounds. The likelihood ratio for surgery was calculated for each horse in the validation data set, using the logistic regression equation. Using Bayes theorem, the posttest probability was calculated, using the likelihood ratio as the test odds and the prevalence of surgery cases (at each institution) as an estimate of the pretest odds. A Hosmer-Lemeshow goodness-of-fit X2 statistic indicated that the model fit the validation data set poorly, as demonstrated by the large X2 value of 26.7 (P < 0.001). However, when the expected proportion of surgical cases was compared with the observed proportion of surgical cases in each of 10 increments of risk, the model's performance appeared satisfactory.